引導式測驗 · 中等 等級
LibriSpeech 語音辨識基準測驗
測試 LibriSpeech 有聲書的起源、朗讀語音域限制、分割獨立性、WER 標準化、解碼器公開和現場測試。
問題 1 的 8
What kind of source material underlies LibriSpeech?
預覽每個問題8 問題
當您參加測驗時,就會出現答案和解釋。
- What kind of source material underlies LibriSpeech?
- Multilingual private phone calls
- Only synthetic voices from one model
- Unscripted classroom discussions
- Read English public-domain audiobooks and text
- Why does a low LibriSpeech WER not prove call-center performance?
- Conversation and telephone acoustics can differ from read audiobooks
- WER cannot be computed on calls
- Every call shares the audiobook narrators
- Read and telephone speech have identical acoustics
- What should a paper name when reporting a LibriSpeech score?
- Only a model nickname
- Only total book count
- The evaluated split and text-normalization/decoder settings
- Only the paper’s abstract
- Why check external pretraining data for overlap?
- Overlap always decreases WER
- A test recording seen during pretraining can compromise independence
- External corpora cannot include books
- A public split name prevents all pretraining overlap
- Which additional data best tests a live-captioning product?
- Consent-based audio from its intended speakers and environments
- Only the benchmark training books
- Only clean generated tones
- Only a dataset card with no audio
- What can change reported WER without changing acoustic model weights?
- Only output font or display color
- Only README length
- Different text normalization or external LM decoding
- Only hardware casing color
- Why measure latency as well as WER for live captions?
- WER already includes display timing
- A correct final transcript may appear too late to help
- All devices have the same delay
- Caption latency cannot be measured
- Which conclusion fits LibriSpeech’s research role?
- It certifies every accent equally
- It replaces deployment validation
- It guarantees speaker identification
- It is a reproducible read-speech reference, not a universal ASR certificate